Why this is in my collection
From the publisher:
Large language models (LLMs) and diffusion models such as ChatGPT and Stable Diffusion have unprecedented potential. With this book, you'll gain a solid foundation in generative AI, including how to apply these models in practice. When first integrating LLMs and diffusion models into their workflows, most developers struggle to coax reliable enough results from them to use in automated systems. Authors James Phoenix and Mike Taylor show you how a set of principles called prompt engineering can enable you to work effectively with AI.
Highlights
- Reliability is the real problem, principles are the answer — Phoenix and Taylor start from the developer's pain that raw LLM output is too erratic for automated systems, and respond with a principled discipline rather than tricks; deterministic scaffolding around a stochastic engine is the structure-beats-magic thesis applied to AI itself.
- Five principles instead of a thousand hacks — give direction, specify format, provide examples, evaluate quality, divide labor: a small, named rule-set that generalizes across models and modalities; compressing a chaotic practice into a compact pattern language is exactly what makes a craft teachable and toolable.
- Specify the output format explicitly — demanding structured output (JSON, schemas, fixed fields) is what turns a text generator into a system component; the model's response becomes data with a contract, which is the only form automation can safely consume.
- Examples are executable specification — few-shot prompting shows the model the pattern rather than describing it; a working demonstration that pattern density in the input governs pattern fidelity in the output, for LLMs as for warehouse generators.
- Evaluate quality, don't vibe it — the principle of building evaluation into the workflow moves prompting from anecdote to measurement; rules and checks around AI output are the forgotten third pillar showing up in the GenAI stack.
- Divide labor: decompose the task into chained steps — big fuzzy requests fail where pipelines of small, verifiable steps succeed; the same decomposition logic that structures a data pipeline or a layered CTE, now applied to cognition.
- Prompts belong in version control — treating prompts as engineering artifacts with versions, tests, and reviews rather than chat improvisations; the moment prompting joined the rules-library rather than the magic drawer.
Highlights on this page are generated with the help of AI.